The Era of Manual Low-Level Optimization is Fading
The era of manual low-level code optimization is coming to an end, hit by a structural pivot: AI agents have begun a full-scale expansion into the inner sanctum of R&D. ByteDance is already deploying specialized agents to write and tune CUDA code, systematically pulling the rug out from under rare and exorbitantly expensive GPU architecture engineers. This isn't just another layer of application software automation; it is a takeover of the fundamental layer where hardware meets code.
Planning Horizons are Becoming Obsolete
The speed of this shift is forcing experts to rewrite their forecasts on the fly. Ajeya Cotra, a leading analyst of AI trajectories, admitted that her January expectations for 2026 are already hopelessly outdated. While Cotra previously estimated that agents would handle 24-hour task planning horizons by the end of 2026, fresh METR benchmarks show that models like a hypothetical 'Opus 4.6' are already managing 12-hour cycles today.
According to Cotra’s revised estimates, agents will consume tasks lasting over 100 hours—equivalent to several full work weeks—by the end of this year. The concept of a human-led "planning horizon" is evaporating as neural networks begin to colonize the very core of the development process.
AI Research & Development Automation (AIRDA)
The strategic focus is shifting toward what researchers from Oxford and GovAI call AIRDA—AI Research & Development Automation. Instead of building chatbots to order pizza, companies are creating the infrastructure for AI to reproduce itself. This recursive cycle allows ByteDance and other major players to bypass the talent shortage. The economic impact is clear:
Automating chip-specific optimization radically shifts infrastructure margins. The release cycle for custom models is significantly shortened. Dependence on scarce systems programmers is neutralized.
We are witnessing a phase where AI is beginning to build itself. If your technical strategy still relies on endless hiring to scale, you have already lost. The priority is no longer just implementing neural networks, but aggressively automating your own R&D engineering layers. Those who continue to polish code by hand risk ending up in the history books alongside punch cards.